[Scalar] Dedicated pass for identifying redundant operations on packed bytes

llvm/include/llvm/InitializePasses.h

@@ -237,6 +237,7 @@ initializeOptimizationRemarkEmitterWrapperPassPass(PassRegistry &);
 LLVM_ABI void initializeOptimizePHIsLegacyPass(PassRegistry &);
 LLVM_ABI void initializePEILegacyPass(PassRegistry &);
 LLVM_ABI void initializePHIEliminationPass(PassRegistry &);
+LLVM_ABI void initializePackedIntegerCombineLegacyPassPass(PassRegistry &);
 LLVM_ABI void initializePartiallyInlineLibCallsLegacyPassPass(PassRegistry &);
 LLVM_ABI void initializePatchableFunctionLegacyPass(PassRegistry &);
 LLVM_ABI void initializePeepholeOptimizerLegacyPass(PassRegistry &);

llvm/include/llvm/Transforms/Scalar.h

@@ -154,6 +154,13 @@ LLVM_ABI FunctionPass *
 createInferAddressSpacesPass(unsigned AddressSpace = ~0u);
 LLVM_ABI extern char &InferAddressSpacesID;
 
+//===----------------------------------------------------------------------===//
+//
+// PackedIntegerCombinePass - Tracks individual bytes through instructions to
+// systematically identify redundant byte packing or unpacking operations.
+//
+LLVM_ABI FunctionPass *createPackedIntegerCombinePass();
+
 //===----------------------------------------------------------------------===//
 //
 // PartiallyInlineLibCalls - Tries to inline the fast path of library

llvm/include/llvm/Transforms/Scalar/PackedIntegerCombinePass.h

@@ -0,0 +1,32 @@
+//===- PackedIntegerCombinePass.h -------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+/// \file
+///
+/// This file provides the interface for LLVM's Packed Integer Combine pass.
+/// This pass tries to treat integers as packed chunks of individual bytes,
+/// and leverage this to coalesce needlessly fragmented
+/// computations.
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TRANSFORMS_SCALAR_PACKEDINTCOMBINE_H
+#define LLVM_TRANSFORMS_SCALAR_PACKEDINTCOMBINE_H
+
+#include "llvm/IR/PassManager.h"
+
+namespace llvm {
+
+class PackedIntegerCombinePass
+    : public PassInfoMixin<PackedIntegerCombinePass> {
+public:
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
+} // end namespace llvm
+
+#endif // LLVM_TRANSFORMS_SCALAR_PACKEDINTCOMBINE_H

llvm/lib/Passes/PassBuilder.cpp

@@ -313,6 +313,7 @@
 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h"
 #include "llvm/Transforms/Scalar/NaryReassociate.h"
 #include "llvm/Transforms/Scalar/NewGVN.h"
+#include "llvm/Transforms/Scalar/PackedIntegerCombinePass.h"
 #include "llvm/Transforms/Scalar/PartiallyInlineLibCalls.h"
 #include "llvm/Transforms/Scalar/PlaceSafepoints.h"
 #include "llvm/Transforms/Scalar/Reassociate.h"

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -121,6 +121,7 @@
 #include "llvm/Transforms/Scalar/MemCpyOptimizer.h"
 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h"
 #include "llvm/Transforms/Scalar/NewGVN.h"
+#include "llvm/Transforms/Scalar/PackedIntegerCombinePass.h"
 #include "llvm/Transforms/Scalar/Reassociate.h"
 #include "llvm/Transforms/Scalar/SCCP.h"
 #include "llvm/Transforms/Scalar/SROA.h"
@@ -542,6 +543,9 @@ PassBuilder::buildO1FunctionSimplificationPipeline(OptimizationLevel Level,
   // opportunities that creates).
   FPM.addPass(BDCEPass());
 
+  // Simplify bit-packed operations before cleaning up with instcombine.
+  FPM.addPass(PackedIntegerCombinePass());
+
   // Run instcombine after redundancy and dead bit elimination to exploit
   // opportunities opened up by them.
   FPM.addPass(InstCombinePass());
@@ -743,6 +747,9 @@ PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
   // opportunities that creates).
   FPM.addPass(BDCEPass());
 
+  // Simplify bit-packed operations before cleaning up with instcombine.
+  FPM.addPass(PackedIntegerCombinePass());
+
   // Run instcombine after redundancy and dead bit elimination to exploit
   // opportunities opened up by them.
   FPM.addPass(InstCombinePass());

llvm/lib/Passes/PassRegistry.def

@@ -476,6 +476,7 @@ FUNCTION_PASS("objc-arc", ObjCARCOptPass())
 FUNCTION_PASS("objc-arc-contract", ObjCARCContractPass())
 FUNCTION_PASS("objc-arc-expand", ObjCARCExpandPass())
 FUNCTION_PASS("pa-eval", PAEvalPass())
+FUNCTION_PASS("packedintcombine", PackedIntegerCombinePass())
 FUNCTION_PASS("partially-inline-libcalls", PartiallyInlineLibCallsPass())
 FUNCTION_PASS("pgo-memop-opt", PGOMemOPSizeOpt())
 FUNCTION_PASS("place-safepoints", PlaceSafepointsPass())

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -104,6 +104,7 @@
 #include "llvm/Transforms/Scalar/InferAddressSpaces.h"
 #include "llvm/Transforms/Scalar/LoopDataPrefetch.h"
 #include "llvm/Transforms/Scalar/NaryReassociate.h"
+#include "llvm/Transforms/Scalar/PackedIntegerCombinePass.h"
 #include "llvm/Transforms/Scalar/SeparateConstOffsetFromGEP.h"
 #include "llvm/Transforms/Scalar/Sink.h"
 #include "llvm/Transforms/Scalar/StraightLineStrengthReduce.h"
@@ -1378,8 +1379,11 @@ void AMDGPUPassConfig::addCodeGenPrepare() {
 
   TargetPassConfig::addCodeGenPrepare();
 
-  if (isPassEnabled(EnableLoadStoreVectorizer))
+  if (isPassEnabled(EnableLoadStoreVectorizer)) {
     addPass(createLoadStoreVectorizerPass());
+    // LSV pass opens up more opportunities for packed integer combining.
+    addPass(createPackedIntegerCombinePass());
+  }
 
   // LowerSwitch pass may introduce unreachable blocks that can
   // cause unexpected behavior for subsequent passes. Placing it
@@ -2101,8 +2105,11 @@ void AMDGPUCodeGenPassBuilder::addCodeGenPrepare(AddIRPass &addPass) const {
 
   Base::addCodeGenPrepare(addPass);
 
-  if (isPassEnabled(EnableLoadStoreVectorizer))
+  if (isPassEnabled(EnableLoadStoreVectorizer)) {
     addPass(LoadStoreVectorizerPass());
+    // LSV pass opens up more opportunities for packed integer combining.
+    addPass(PackedIntegerCombinePass());
+  }
 
   // LowerSwitch pass may introduce unreachable blocks that can cause unexpected
   // behavior for subsequent passes. Placing it here seems better that these

llvm/lib/Transforms/Scalar/CMakeLists.txt

@@ -61,6 +61,7 @@ add_llvm_component_library(LLVMScalarOpts
   NaryReassociate.cpp
   NewGVN.cpp
   PartiallyInlineLibCalls.cpp
+  PackedIntegerCombinePass.cpp
   PlaceSafepoints.cpp
   Reassociate.cpp
   Reg2Mem.cpp